sys: Use DXCore for accelerator discovery

[xieofxie]

Summary

• use Windows DXCore as the primary source of GPU/NPU adapter identity in winml sys
• keep WMI/PnP as descriptive enrichment for driver and manufacturer, and as a compatibility fallback when DXCore is unavailable
• preserve one native row per physical adapter, so identical display names no longer require WMI-to-ORT matching
• surface the DXCore adapter LUID in text, compact, and JSON output as a side product of native system discovery

Discovery behavior

DXCore returns the adapter name, hardware IDs, and LUID together. winml sys now creates accelerator rows from that inventory, then joins WMI/PnP metadata by hardware ID or name. ONNX Runtime remains responsible only for execution-provider inventory and EP-specific facts; it no longer assigns the top-level device LUID.

CPU discovery remains WMI-based and reports LUID: N/A because CPU devices do not have a Windows graphics/compute adapter LUID.

Hardware observations

Feature	Selected configuration	Discovered device	LUID
TensorRT RTX GPU	--ep nvtensorrtrtx --device gpu	NVIDIA GeForce RTX 4080	0x00000000_0x000175F5
QNN GPU	--ep qnn --device gpu	Qualcomm(R) Adreno(TM) X1-85 GPU	0x00000000_0x00012AC1
QNN NPU	--ep qnn --device npu	Snapdragon(R) X Elite - X1E80100 - Qualcomm(R) Hexagon(TM) NPU	0x00000000_0x000135AA

All three accelerator paths were revalidated through DXCore on this revision. QNN GPU inference completed at 0.743 ms mean latency and 1,345.89 samples/sec.

GPU: NVIDIA GeForce RTX 4080 (LUID: 0x00000000_0x000175F5) | CPU: Intel(R) Core(TM) i9-10900X CPU @ 3.70GHz (LUID: N/A)

Validation

• 134 expanded sysinfo/device unit tests passed
• 2 hardware E2E device-shape and cross-view tests passed
• direct DXCore probe returned the physical NVIDIA GeForce RTX 4080 and filtered the software Basic Render Driver
• QNN NPU model selection succeeded
• required mypy check passed for 440 source files
• Ruff clean on changed Python files

[Zhenchao Ni (zhenchaoni)]

I found two device-to-LUID matching issues that can produce missing or incorrect identifiers.

[xieofxie]

The review discussion exposed that top-level LUID assignment should not depend on ORT-to-WMI name matching. Commit 6563659 replaces that design: DXCore now owns each GPU/NPU row and returns its name, hardware IDs, and LUID together; WMI/PnP only enriches descriptive fields. This also handles multiple physical adapters with identical display names because each DXCore row already carries its own identity.

[xieofxie]

Latest Qualcomm retest confirms the native path: DXCore returned Adreno X1-85 (0x00000000_0x00012AC1) and Snapdragon X Elite Hexagon NPU (0x00000000_0x000135AA). QNN GPU inference measured 0.743 ms mean latency / 1,345.89 samples/sec; QNN NPU selection succeeded. The prior lint failure was a mypy-only ctypes annotation issue, fixed in a2ab36e; mypy -p winml.modelkit now passes all 440 source files.